A Topological Framework for Local Structure Analysis in Condensed Matter

Emanuel A. Lazar1 , Jian Han1 , David J. Srolovitz1,2

1
Department of Materials Science and Engineering
Department of Mechanical Engineering and Applied Mechanics
University of Pennsylvania, Philadelphia, Pennsylvania 19104, USA.
(Dated: August 26, 2015)

arXiv:1508.05937v1 [cond-mat.other] 24 Aug 2015

Physical systems are frequently modeled as sets of points in space, each representing the position
of an atom, molecule, or mesoscale particle. As many properties of such systems depend on the
underlying ordering of their constituent particles, understanding that structure is a primary objective
of condensed matter research. Although perfect crystals are fully described by a set of translation
and basis vectors, real-world materials are never perfect, as thermal vibrations and defects introduce
significant deviation from ideal order. Meanwhile, liquids and glasses present yet more complexity.
A complete understanding of structure thus remains a central, open problem. Here we propose
a unified mathematical framework, based on the topology of the Voronoi cell of a particle, for
classifying local structure in ordered and disordered systems that is powerful and practical. We
explain the underlying reason why this topological description of local structure is better suited for
structural analysis than continuous descriptions. We demonstrate the connection of this approach
to the behavior of physical systems and explore how crystalline structure is compromised at elevated
temperatures. We also illustrate potential applications to identifying defects in plastically deformed
polycrystals at high temperatures, automating analysis of complex structures, and characterizing
general disordered systems.

Condensed matter systems are often abstracted as

large sets of points in space, each representing the position of an atom, molecule, or mesoscale particle. Two
challenges frequently encountered when studying systems
at this scale are classifying and identifying local structure. Simulation studies of nucleation, crystallization,
and melting, for example, as well as those of defect migration and transformation, require a precise understanding of which particles are associated with which phases,
and which are associated with defects. As these systems
are abstracted as large point sets, these dual challenges
of classifying and identifying structure reduce to ones of
understanding arrangements of points in space.
A primary difficulty in classifying structure in spatial
point sets arises from a tension between a desire for completeness and the necessity for practicality. The local
neighborhood of a particle within an ensemble of particles can be completely described by a list of relative
positions of each of its neighbors. However, while such
a list of coordinates is complete in some sense, this raw
data provides little direct insight, leaving us wanting for a
practical and more illuminating description. This tension
is often mediated by the choice of an order parameter,
which distills structural data into a single number or vector, and which is constructed to be both informative and
computationally tractable [1].
A central limitation of conventional order parameters
is exhibited in degeneracies that arise in describing neighborhoods that are structurally distinct but which map
to identical order-parameter values. Some order parameters classify particles in face-centered cubic (FCC) and
body-centered cubic (BCC) crystals identically, while
others classify particles in FCC and hexagonal closepacked (HCP) crystals identically [2]. Similarly, particles located near defects in a low-temperature crystal can

have order-parameter values identical to those of particles in a high-temperature defect-free crystal. These degeneracies point to an inherent incompleteness in such
order-parameter classifications of local structure. Consequently, different order parameters are necessary to study
different systems [1, 2].

FIG. 1. The frame of a Voronoi cell of a central particle (blue),

surrounded by its nearest neighbors (gold). The topology of
the Voronoi cell captures structural information about the
local neighborhood.

In this paper we propose a mathematical framework

to classify local structure that avoids much of the degeneracy encountered in other approaches and which, therefore, is equally applicable to all ordered and disordered
systems of particles. More specifically, the local structure around a particle is classified using the topology of
its Voronoi cell (see Fig. 1). Families of Voronoi cell
topologies are constructed by considering those topologies that can result from infinitesimal perturbations of
an ideal structure. We demonstrate that this classification scheme is consistent with the manner in which local
ordering changes in high-temperature single crystals as
the temperature is raised toward their melting points.
We highlight a potential use of this approach for visualizing defects in crystalline solids at high temperatures,

2
and contrast it with previous methods. We then demonstrate an application of this approach to the automated
analysis of the evolution of complex structures, where
conventional methods are often inadequate. Finally, we
show an application in which this approach is used to
provide robust statistical-structural descriptors for characterizing disordered systems.

I.
A.

THEORY

The Configuration Space of Local Structure

A deeper understanding of local structure can be developed through consideration of all possible arrangements
of neighbors of a central particle. The local neighborhood
of a particle within an ensemble of particles can be completely described by a vector of relative positions of its
n nearest neighbors: x = (r1 , r2 , ..., rn ), where ri is the
relative position of the ith neighbor of a central particle.
For suitably large n, any question about the local neighborhood of a particle can be answered through complete
knowledge of x. We use C(n) to denote the configuration
space of all possible arrangements of n nearest neighbors:
C(n) = {(r1 , r2 , ..., rn ) : ri R3 }.

(1)

Each point in C(n) thus corresponds to a specific local arrangements of particles. Figure 2(d) provides a schematic
of C(n) and highlights points corresponding to local arrangements of BCC, FCC, HCP, and diamond structures.
As defined in Eq. 1, the dimension of C(n) is 3n; ignoring
rotations and scaling reduces the dimension of C(n) by 4.
Ignoring permutations of the n neighbors and disallowing
multiplicities further changes the geometry and topology
of C(n), but not its dimension.
Order parameters can be thought of as functions that
map C(n) to a lower-dimensional order-parameter space;
order-parameter spaces most commonly used are Rd ,
where d is substantially smaller than 3n. Each choice
of order parameter results in a different subdivision of
C(n) into regions on which that order parameter is constant; for real-valued continuous functions, these regions
are commonly known as level sets. To help understand
the degeneracy observed in continuous order-parameter
methods, consider that for every continuous mapping
from an unbounded high-dimensional space to a lowerdimensional space, there exist points x1 , x2 arbitrarily
far apart, but for which (x1 ) and (x2 ) are identical
[3]. The continuity of an order parameter thus entails
the kind of degeneracy highlighted above. In contrast,
discrete order parameters are not subject to this limitation, as distances between points with identical orderparameter values can be bounded. This motivates the
question of how to reasonably subdivide C(n). We now
show that Voronoi topology offers one such approach.

FIG. 2. (a-c) Voronoi cells of particles in BCC, FCC, and

HCP crystals. Vertices at which more than four Voronoi cells
meet are marked by red circles. Small perturbations of the
particle positions result in topological changes near these vertices. (d) Schematic of C(n), the space of all possible configurations of n neighbors. This space can be divided into regions
on which the Voronoi cell topology is constant. The topology
of a point that lies on the boundaries of multiple regions is
unstable, and infinitesimal perturbations of the neighbors will
result in a change of topology. The inset shows the neighborhood around xFCC ; unshaded regions indicate primary types,
while shaded ones indicate secondary types.

B.

Voronoi Topology

For a fixed set of particles, the Voronoi cell of a central

particle is the region of real space closer to that particle than to any other [4]. Figure 1 illustrates a central
particle, its Voronoi cell, and fifteen neighboring particles. Two particles whose Voronoi cells share a face are
called neighbors. We identify two Voronoi cells as having
the same topology, or topological type, if there exists
a one-to-one correspondence between their sets of faces
that preserves adjacency.
The topology of the Voronoi cell of a particle describes
the manner in which neighbors of a particle are arranged
relative to it. An n-sided face, for example, indicates a
pair of particles which have n neighbors in common. The
topology of a Voronoi cell thus provides a robust description of how neighbors are arranged not only relative to a
central particle, but also to one another. In this sense, it
is a good description of local structure.
Voronoi cell topology also provides a natural decomposition of C(n) into regions in which the Voronoi cell
topology is constant, as illustrated in Fig. 2(d). We consider this decomposition natural because it allows us to
coarse-grain the effects of small perturbations on local
structure. Small perturbations of the particle coordinates correspond to small displacements in C(n), and
since the Voronoi topology does not change for almost all
points under small perturbations, these small perturbations, which are often unimportant, are naturally ignored
without the introduction of an artificial cut-off [2].

3
Voronoi cell topology was first introduced by Bernal
and others to study the atomic structure of liquids [57],
and has been subsequently applied to study a wide range
of condensed matter systems, including random sphere
packings [7, 8], finite-temperature crystals [9], and metallic glasses [10]. In those studies, however, the topology
of a cell was characterized by counting its types of faces
(e.g., triangles and quadrilaterals), though it ignored the
way in which those faces are arranged. While this limited
description has been used to study some aspects of crystallization [11], it cannot distinguish particles whose local
environments are FCC from those whose local environments are HCP, as both Voronoi cells have twelve foursided faces. In previous work [12, 13], the authors have
shown how to use a graph-tracing algorithm introduced
by Weinberg [14] to efficiently compute strings which encode a complete description of the Voronoi cell topology;
see Methods for further details.
A second limitation arising in traditional Voronoi approaches results from abrupt changes in topology due to
small geometric perturbations. Consider, for example,
that Voronoi cells of particles in FCC and HCP crystals are topologically unstable since some vertices are
shared by more than four Voronoi cells (see Fig. 2), infinitesimal perturbations of the particle positions, such
as those arising from non-hydrostatic strain or thermal
vibrations, will change their topology [15]. This problem has been sufficiently challenging to limit the utility of conventional Voronoi approaches in studying even
slightly perturbed crystal structures [2]. This problem
can be solved through the classification of topological
types described in the following section.

C.

Theory of -types

In this section we show how topological types can be

classified using the approach developed in the previous
two sections. On a basic level, every arrangement of
neighbors relative to a central particle can be described
by its Voronoi cell topology. Families of topological types
associated with a particular structure can then be defined
as sets of types obtained through infinitesimal perturbations of that structure. This classification scheme enables
a description of the effects of small strains and thermal vibrations on local structure, and provides a robust framework suitable for theoretical and numerical analysis.
Every local arrangement of neighbors is described by
a distinct point x in C(n), and subsequently corresponds
to a unique Voronoi cell topology V [x ]. For example,
if = BCC, then x = xBCC describes a particle that
has the same local environment as a particle in a perfect
BCC crystal; its Voronoi cell topology V [xBCC ] is the
truncated octahedron, illustrated in Fig. 2(a).
A suitable distance function on C(n) allows us to define sets of topological types associated with infinitesimal
perturbations as follows. We let B (x) be a ball of radius  centered at x. This region of C(n) corresponds to

configurations obtained through small perturbations of a

particle and its neighbors, where  controls the magnitude of such a perturbation. The set of topological types
obtained from all possible perturbations of x with magnitude smaller than  is denoted V [B (x)]. We define the
family of topological types associated with infinitesimal
perturbations of as the limiting set:
F := lim V [B (x )] .
0

(2)

In more physical terms, F is the set of all topological

types that can be obtained through arbitrarily small perturbations of a central particle and its neighbors. The
Voronoi cell topology of points in the interior of a region in C(n) remains unchanged by small perturbations.
In contrast, points such as xFCC and xHCP are located
at the boundaries of multiple regions, and small perturbations entail a change in Voronoi cell topology. Thus,
FFCC and FHCP consist of multiple topological types,
whereas FBCC , located in the interior of a region, consists of a single type. If a topological type is in F , then
we say that it is a -type. Note that a topological type
can belong to multiple families; this indeterminacy will
be considered below. This classification of -types allows
us to account for topological instability without modifying the sample data by collapsing edges or faces using ad
hoc criteria (e.g., cut-offs) [2, 9].
Among -types, a further distinction can be drawn
based on the manner in which the Voronoi cell topology subdivides C(n). Using a suitable volume measure
Vol, we define the ideal frequency f ( ) of a topological
type relative to x as follows:


Vol V 1 [ ] B (x )
f ( ) := lim
,
(3)
0
Vol (B (x ))
where V 1 [ ] is the set of points in C(n) whose Voronoi
cell have topology . If f ( ) > 0, we call a primary
-type; if f ( ) = 0, we call it a secondary -type. The
inset in Fig. 2(d) highlights a number of regions incident
with xFCC . Some of those regions meet xFCC at finite
solid angles; therefore, their fractional volumes within
an -ball converge to positive values as  0; these are
primary FCC-types. In contrast, fractional volumes tend
to zero as  0 for other regions which meet xFCC at
cusps; these are secondary FCC-types.
The distinction between primary and secondary types
appears to result from the manner in which topological
instabilities resolve when perturbed. To illustrate this
distinction, Fig. 3(a) shows an unstable vertex shared by
six Voronoi cells in FCC or HCP crystals; such vertices
are marked by red circles in Figs. 2(b) and (c). Figure 3(b) depicts the most common manner in which such
an unstable vertex resolves upon small perturbations of
neighboring particles [15]. In this resolution, a new foursided face is formed between two non-adjacent Voronoi
cells; all unstable vertices resolve in this manner in primary types. A less common resolution can also occur as
a result of cooperative motion of neighboring particles.

FIG. 3. (a) An unstable vertex shared by six Voronoi cells in

FCC or HCP crystals; such vertices are marked by red circles
in Figs. 2(b) and (c). A small perturbation will cause the
vertex to resolve into either (b) a four-sided face, or (c) a pair
of adjacent triangular faces; these resolutions are associated
with primary and secondary types, respectively.

In this resolution, depicted in Fig. 3(c), two triangular

faces are created [11]; secondary types can include such
resolutions.
Determining F is feasible through consideration of
all possible ways in which unstable vertices can resolve.
For example, the ideal FCC Voronoi cell, illustrated in
Fig. 2(b), has six unstable vertices. In primary types,
each such vertex resolves in a manner illustrated in
Fig. 3(b), in one of three directions. More specifically,
the unstable vertex can transform in such a way that the
central cell gains a square face, or else gains an edge in
one of two directions. We consider all possible combinations of these resolutions over the six unstable vertices,
and calculate the topological types of the resulting polyhedra using the algorithm developed in [13]; a total of
44 distinct topological types occur in this manner. In
secondary types, unstable vertices can also resolve in the
manner illustrated in Fig. 3(c), or else remain unstable.
An additional 6250 topological types can occur in this
manner. A similar approach can be followed to determine F for other structures. Additional details can be
found in the supplementary material.

II.

FINITE-TEMPERATURE CRYSTALS

The proposed distinction between primary and secondary types is supported by atomistic simulation. We
studied the atomic structure of three model materials,
BCC tungsten [16], FCC copper [17], and HCP magnesium [18], at elevated temperatures using molecular
dynamics (MD) in the N P T ensemble [19]. Simulated
systems contained 1,024,000, 1,372,000, and 1,029,600
atoms, respectively, in periodic supercells. In each simulation, a defect-free crystal was heated from T = 0 in
increments of 50 K and equilibrated for 50 ps at each
temperature. Figure 4 shows how the distribution of
topological types changes with temperature; each curve
indicates the frequency of a single topological type.
Types can be grouped according to the shape of their
frequency curves. Frequencies of one group of types approach finite values as T 0, and change very little with
temperature (blue curves in Fig. 4). Frequencies of a second group rapidly approach zero as T 0 (pink curves).

Types of a third group only appear at high temperatures

(grey curves). Remarkably, these groups correspond exactly to the sets of primary , secondary , and non-types for each system, as enumerated using the analysis
of instabilities, described above. The theory of -types is
thus consistent with observed results and suggests a new
approach for analyzing thermal effects.
One noteworthy feature of Fig. 4 is the similarity between FCC and HCP, in contrast to BCC. These general behaviors appear to depend primarily on the crystal structure rather than on bonding particulars. Indeed, preliminary investigations show that when atoms
in BCC, FCC, and HCP crystals are perturbed from their
equilibrium positions with a Gaussian noise whose magnitude scales with temperature (i.e., an Einstein model
[20]), the distribution of topological types changes in a
manner similar to that observed in Fig. 4.
A second notable feature is the total frequency of types in the liquid phases of the three systems. In liquid
tungsten, the unique BCC-type accounts for less than
0.05% of all atoms just above Tm , where Tm is the bulk
melting temperature. In contrast, liquid copper consists
of roughly 20% FCC-types, and liquid magnesium consists of roughly 30% HCP-types, just above Tm . These
structural data are relevant in studying physical processes such as crystallization [21] and melting [22].
A third feature that stands out is the high fraction of
-types in the FCC and HCP samples at temperatures
just below melting. In FCC, 94% of all atoms are classified as having FCC local structure just below melting; in
HCP this number is 96%. At 0.85Tm , these numbers are
over 99% in both systems. This suggests a natural use
of -types for identifying structure in highly-perturbed
atomistic systems, such as those at high temperatures.
We next consider several applications of the proposed
approach to illustrate some of its unique features.

III.

DEFECT VISUALIZATION

As noted, the high frequencies of -types in single crystals, even at extremely high temperatures, suggests their
use for visualization of local structure in atomic systems.
Figure 5 shows thin cross-sections from an FCC aluminum polycrystal prepared using MD [23]. The sample was obtained by annealing a microstructure obtained
through simulated grain-growth [24], plastically deforming it at low temperature, and then thermalizing it at
0.9Tm . In these figures, atoms that are FCC-types are not
shown for clarity; among the remaining atoms, those that
are HCP-types are shown in gold, and all other atoms are
shown in dark blue.
In Fig. 5(a), grain boundaries can be identified as a
network of non-FCC-type atoms. Vacancies (A) can be
identified within the grain interiors as small clusters of
non-FCC-types; as only a thin cross-section of the material is shown, not all atoms adjacent to these defects
appear in the figure. A twin boundary (B) and stack-

5
100

Primary

(a) BCC tungsten

Primary
Secondary

(b) FCC copper

Primary
Secondary

(c) HCP magnesium

Frequency

10-1
10-2
10-3
10-4
10-5 0

1000

2000

3000

4000

5000

Temperature (K)

6000

500

1000

1500

Temperature (K)

2000

200

400

600

800

1000

Temperature (K)

1200

1400

FIG. 4. Frequencies of all -types and all non--types that appear in the single-crystal (a) BCC tungsten, (b) FCC copper, and
(c) HCP magnesium as a function of temperature upon heating from T = 0 to 150% of the bulk melting temperature. Blue
curves indicate primary -types, pink curves secondary -types, and grey curves non--types. Thick curves indicate the sum
of all frequencies of the corresponding color; note that there is only one primary -type and no secondary -types for BCC.

ing fault (C) can be identified as one and two layers of

gold-colored (HCP, non-FCC type) atoms, respectively.
Figures 5(b) and (c) show magnified images of a dislocation and a stacking fault inclined at a low angle relative
to the cross-section plane.

As noted earlier, individual Voronoi topologies can belong to multiple families; we use the term indeterminate type to refer to such cases. This indeterminacy
complicates the visualization procedure suggested here,
as many types in FHCP also belong to FFCC . For this reason some atoms that belong to the twin boundaries and
stacking faults are not seen in Figs. 5(a) and (c). This
shortcoming can be easily addressed within the topological framework, and is discussed in the supplementary
material.

The utility of the proposed topological framework for

local structure classification and identification is useful
for finite-temperature simulations of atomic systems containing defects. Of particular interest are the many phenomena which are only of interest at high temperature,
such as dislocation climb [25], interface thermal fluctuation [26], and defect kinetics under irradiation conditions
[27]. Conventional visualization methods require quenching or temporally averaging a sample prior to crystal defect analysis [2]. In general, we do not know whether
this tampering with the data leads to significant discrepancies between the observed structures and the actual finite-temperature ones. Remarkably, the proposed
approach identifies all defects in Fig. 5 and does not erroneously identify any bulk atoms as being adjacent to
defects, all at very high temperature and without quenching or time averaging. The topological approach thus
provides a natural method for identifying and visualizing
local structure that involves no ad hoc cut-off parameters
and which is robust at high temperatures. This opens
new opportunity for in situ structure analysis of atomic
simulations at temperature, potentially identifying new
high-temperature mechanisms and defect structures.

IV.

COMPARISON WITH OTHER METHODS

Although a complete survey of existing methods for

analyzing structure in high-temperature atomic systems
is beyond the scope of this introductory paper, we briefly
consider how visualization using -types compares with
three of the most frequently-used order parameters: centrosymmetry [28], bond-angle analysis [29], and adaptive
common-neighbor analysis [2, 30].
As a concrete example, we consider a stacking-fault
tetrahedron [31] in a high-temperature FCC aluminum
single crystal. A stacking-fault tetrahedron (SFT) is a
three-dimensional defect consisting of four stacking faults
that form the faces of a tetrahedron. The interior of an
SFT is an FCC crystal; its edges are stair-rod dislocations [25]. Figure 6 illustrates a cross-section through
the center of an SFT and parallel to one of its four faces;
the intersection of the SFT boundary with the viewing plane is an equilateral triangle. This perfect SFT
was constructed in FCC copper and then thermalized
at 0.85Tm . Centrosymmetry, bond-angle, and adaptive
common-neighbor analyses were all performed using the
OVITO software package [32].
Figure 6(a) shows atoms colored using the centrosymmetry order parameter. In this coloring, atoms belonging to faces of the SFT have higher centrosymmetry values than those in the FCC environment, as expected.
Note, however, that many atoms inside and outside the
SFT also have high centrosymmetry values. Such atomic
environments do not, however, indicate crystal defects,
but rather result from thermal fluctuations which locally
distort the atomic environment. The inability of the
centrosymmetry order parameter to distinguish structural defects from thermal perturbations requires users
to quench a system before analyzing it.
Figures 6(b) and (c) show atoms colored using bondangle analysis and adaptive common-neighbor analysis,
respectively, also at 0.85Tm . In these figures, many atoms
belonging to the SFT faces are classified as having HCP
local structure, as expected. However, both methods
identify many atoms away from the stacking faults as

(a) Polycrystal cross-section

(b) Dislocation

(c) Stacking fault

FIG. 5. Polycrystalline aluminum at 938 K (0.9Tm ); the width

of each cross-section is 2 nm. Atoms that are FCC-types are
not shown for clarity. Of the ones remaining, those that are
HCP-types are shown in gold and all other atoms are shown
in dark blue. Grain boundaries are seen as a network of nonFCC-types (dark blue and gold atoms). In cross-section (a),
defects are labeled as follow: vacancies A, twin boundary B,
and stacking fault C. Cross-sections (b) and (c) show magnified images of a dislocation and stacking fault.

structural defects, despite the absence of other defects in

the crystal. Moreover, application of bond-angle analysis
incorrectly identifies many atoms in the bulk as having
HCP local structure. Although the general shape of the
SFT can be discerned in both figures, the details are ambiguous and automated location of the SFT in an atomic
ensemble is difficult or impossible at the simulation temperature.
These results are in contrast with the picture produced
using Voronoi topology, illustrated in Figure 6(d). The
approach taken here provides the clearest representation
of the SFT. In this case, every atom characterized as being in an HCP environment is on an SFT face, without
exception, despite the high temperature and the strain
fields of the constituent partial dislocations. Moreover,
all atoms not at the surface of the SFT are correctly

(a) Centrosymmetry

(b) Bond-angle analysis

(c) Adaptive CNA

(d) Voronoi Topology

FIG. 6. Cross-section of a stacking-fault tetrahedron in copper at 85% of its melting temperature, colored using several
popular visualization approaches, and the proposed one. In
(b) and (c), dark blue, yellow and red indicate atoms in FCC,
HCP, and other local environments, respectively. In (d), dark
blue, yellow, and red indicate atoms that are FCC-types,
HCP- but not FCC-types, and all other types.

identified as being FCC-type. Finally, atoms lying at the

corners of the triangular cross-section through the SFT
triangle are identified as neither stacking faults (HCPtype) nor bulk type, but as having a distinct local structure; these are the dislocation cores. The sole weakness
of this visualization procedure results from indeterminate
types which belong to both FFCC and FHCP and whose
neighborhoods are identified as FCC rather than HCP.
This limitation can be addressed and is discussed in the
supplementary material.
This topological approach to structure visualization
can also be applied to low-temperature systems such as
those obtained through quenching (inherent structures);
an example can be found in the supplementary material.

V. GRAIN-BOUNDARY
CHARACTERIZATION AND ANALYSIS

The proposed framework also enables the analysis of

structurally-complex systems in an automated manner.
To illustrate this capability we analyze how a particular
grain boundary transforms between a series of distinct
structures as a result of absorbing point defects, as it
may, for example, under irradiation conditions.
In particular, we consider a 5 [001] (310) symmetric tilt boundary in a BCC tungsten bicrystal. This
grain boundary exhibits three structurally distinct, stable phases. We begin by characterizing these phases

(b) Phase II

(c) Phase III

(d)

FIG. 7. (a-c) Three structurally distinct stable phases of the

5 [001] (310) symmetric tilt boundary in BCC tungsten, arbitrarily labeled Phase I, Phase II, and Phase III. Atoms are
colored according to topological type, each assigned a unique
color; BCC-type atoms are not shown for clarity. (d) An
image from a large simulation of the same boundary, initially
constructed uniformly of the Phase I structure, and into which
self-interstitial atoms have been randomly placed at a constant rate at 1500 K. Atoms are colored by topological type,
as shown in (a-c); BCC-type atoms are not shown for clarity;
all other atoms are shown in grey.

using the Voronoi topologies of the constituent atoms;

these phases are illustrated in Fig. 7(a-c). Atoms are
colored according to their topological type; BCC-type
atoms are not shown for clarity. Phase I consists of three
distinct topological types, colored in different shades of
blue, Phase II consists of two distinct topological types,
colored in shades of green, and Phase III consists of six
topological types, colored in shades of red, orange, and
yellow.
We initialize the simulation by constructing a large
bicrystal containing a 5 [001] (310) symmetric tilt
boundary at 0 K and uniformly of the Phase I structure.
The sample is then heated to 1500 K, and equilibrated at
this temperature for 4 ns. Self-interstitial atoms are then
inserted into random locations in the boundary plane at
a rate of 1.62 atoms/nm2 /ns. We analyze the resulting grain boundary structure throughout the MD simulation using Voronoi topology. Figure 7(d) shows a grain
boundary with distinct domains of all three grain boundary phases, suggesting a phase transition driven by absorption of self-interstitial atoms.
To study the transformation of the grain boundary

Figure 8 shows the fraction of the three phases over

time, starting when the first self-interstitial atom is
added to the grain boundary. During the initial 100 ps,
there is a sharp decrease in Phase I, accompanied by a
rapid growth of Phase III. After approximately 300 ps,
the grain boundary structure settles into a pattern of increasing and decreasing Phase I, Phase II and Phase III
fractions, all with the same period. The minimum in
the Phase III fraction corresponds to the maximum in
the Phase I fraction and the maximum in the Phase II
fraction corresponds to minima in the Phase I and III
fractions. The period is commensurate with the time
required to add a full (310) plane of atoms to the sample. At this temperature, the equilibrium grain boundary
structure is dominated by Phase III, Phase I (the equilibrium structure at 0 K) never completely disappears,
and Phase II is present only in a very small fraction of
the grain boundary.

0.5
Fraction of Grain Boundary

(a) Phase I

structure, we track the fraction of each phase present

during the evolution. We begin by counting the number
of atoms in the sample with topological types associated
with each of the three phases. We next calculate the
number of non-BCC-type atoms per nm2 in each of the
three phases. Finally, we normalize the -type counts for
the three phases by dividing by the number of -types
per unit area and the total grain boundary area.

Phase I
Phase II
Phase III

0.4
0.3
0.2
0.1
0

500

1000

1500 2000 2500

Time (ps)

3000

3500

4000

FIG. 8. The fraction of the 5 [001] (310) symmetric tilt

boundary in BCC tungsten occupied by its structurally distinct, stable phases during the insertion of self-interstitial
atoms at a fixed rate.

This example illustrates the power of the topological

approach for automating structure analysis in systems
with complex defect structures and for determining defect statistics. We defer a more complete analysis of this
example to a future study of grain boundary structure
evolution during irradiation.

8
VI.

DISORDERED STRUCTURES

Finally, we consider how the proposed approach can

be used to characterize disordered systems such as liquids
and glasses. In contrast to conventional order parameters
which are typically useful for studying either ordered or
disordered systems, but not both the approach taken
here can be applied effectively to all kinds of systems.
As the topological type of each Voronoi cell provides a
robust structural description of the local neighborhood
of a particle, the distribution of topological types in a
system can be interpreted as a statistical-topological description of the system as a whole. This ability to characterize both ordered and disordered systems within the
same framework is of particular importance for studying
phase transitions between ordered and disordered phases,
as well as between distinct disordered phases.
Using MD, we simulate two disordered systems of copper atoms: a high-temperature liquid (HTL) equilibrated
at roughly 1.85Tm , and a glass-forming liquid (GFL)
obtained by undercooling the initial liquid to roughly
0.75Tm ; each system contained 1,372,000 atoms. The distributions of topological types in the two systems enable
us to describe structural features of the systems in a robust and quantitative manner, and to observe structural
differences between them.
0.045
0.040

High-temperature liquid
Glass-forming liquid

Frequency

0.035
0.030
0.025
0.020
0.015
0.010
0.005
0

Topological Types

FIG. 9. Frequencies of the twenty-five most common topological types in liquid copper at 1.85Tm , and corresponding
frequencies in the glass-forming liquid copper at 0.75Tm ; circles indicate frequencies in quenched samples.

Bars in Figure 9 show frequencies of the fifty most

common topological types in the HTL, and corresponding frequencies in the GFL. The most common types in
the HTL do not occur nearly as frequently as the most
common types in the GFL. In particular, the sum of frequencies of the five most common topological types in
the GFL is greater than the sum of frequencies of the
fifty most common types in the HTL. In this sense, the
GFL is substantially more ordered than the HTL. Note
that in an ideal gas, in which there is no interaction be-

tween particles, the distribution of types is considerably

less concentrated than in either of these systems [33].
Circles in Figure 9 indicate corresponding frequencies in systems obtained by quenching these two systems. When quenched, frequencies of these common
types change only modestly. This again indicates that
the current method is relatively insensitive to thermal
vibrations.
Other means of quantifying disorder in atomic systems
have been widely developed, and have been used to distinguish distinct types of disordered systems. Structural
correlation functions [1] and configurational entropy [34
36] are both non-local descriptors that have been used to
study disordered systems. Other recent work has focused
on local structural measures [5, 10]; our approach is in
this vein.

VII.

CONCLUSIONS AND DISCUSSION

We have introduced a mathematical approach towards

classifying and identifying local structure of a particle
within a system of particles. Applications highlight its
utility in analyzing atomistic data sets, such as those produced by molecular dynamic simulations. In particular,
the theory of -types enables identification and visualization of defects in ordered systems at high temperatures. This capability can play an important role for
in situ study of high-temperature mechanisms currently
inaccessible to current structure-identification methods.
The proposed framework also enables a new approach
for characterizing and identifying defects. This in turn
allows for an automated approach for studying systems in
which structural features evolve. Finally, Voronoi topology enables the characterization of disordered systems in
a statistical manner, through the distribution of topological types. We have illustrated the potential of this approach in distinguishing a high-temperature liquid from
a glass-forming liquid.
Any description of structure within a fixed distance of
a particle will be unable to capture all long-range structural features of a system. Figure 5 provides clear examples of this limitation, where an atom with HCP local
environment might be part of a twin boundary, stacking fault, or other defect; further analysis is required to
distinguish between these. The analysis of local structure introduced here can be integrated into tools such as
those developed in [37] to automate long-range structural
analysis.
The authors have developed software to automate this
analysis, and is available upon request.

MATERIALS AND METHODS

Deciding whether two Voronoi cells have the same

topological type is equivalent to deciding whether two

9
planar graphs are identical, as the edge-boundary of every Voronoi cell is a planar graph. For each particle in
a system we compute a code that records the graph
structure of the edge-boundary network of its Voronoi
cell. To do this, we first determine the Voronoi cell using the Voro++ software package [38], which computes a
list of faces, each represented as an ordered sequence of
vertices. Next, we use a graph-tracing algorithm to compute a code for this planar graph. More specifically, the
following algorithm of Weinberg [14] is followed: (a) An
initial vertex is chosen and assigned the label 1. (b) An
edge adjacent to that vertex is chosen and travel begins
along that edge. (c) If an unlabeled vertex is reached,
it is labeled with the next unused integer and we turn
right and continue. (d) If a labeled vertex is reached
after traveling along an untraversed edge, we return to
the last vertex along the same edge but in the opposite
direction. (e) If a labeled vertex is reached after traveling
along an edge previously traversed in the opposite direction, we turn right and continue; if that right-turn edge
has also been traversed in that direction, we turn along
the next right-turn edge available; if all outgoing edges
have been traversed, we stop. At this point, each edge in
the graph has been traversed once in each direction; the
ordered list of the vertices visited is called a code.

tex and edge, and for each of two spatial orientations;

all labels are cleared before producing each code. For a
Voronoi cell with e edges, 4e codes are generated, each
an ordered list of 2e integer labels. Each code completely
describes the Voronoi cell topology, and so it is sufficient
to only record one of them. A code for a typical Voronoi
cell requires less than 100 bytes of storage. Additional
details can be found in [12, 14, 39].
Run-time. The use of Voronoi topology for structure
identification is computationally efficient. In preliminary
tests, the Voronoi topology of one millions atoms could
be calculated on a desktop computer in less than one
minute. By contrast, conventional approaches used in
high-temperature structure analysis require that systems
be quenched before visualization. A complete quench
necessary to obtain the inherent structure can require
several hours of computation for a system of comparable
size.

ACKNOWLEDGMENTS

Codes are constructed for each choice of initial ver-

We gratefully acknowledge discussions with and assistance from Chris H. Rycroft and Zhaoxuan Wu. Figures
5, 6, and 7 were created with AtomEye [40]. EAL and
DJS acknowledge support of the NSF Division of Materials Research through Award 1507013.

[1] Truskett, T. M., Torquato, S. & Debenedetti, P. G.

Towards a quantification of disorder in materials: Distinguishing equilibrium and glassy sphere packings.
Phys. Rev. E 62, 993 (2000).
[2] Stukowski, A. Structure identification methods for
atomistic simulations of crystalline materials. Modelling
Simul. Mater. Sci. Eng. 20, 045021 (2012).
[3] Landweber, P. S., Lazar, E. A. & Patel, N. On fiber
diameters of continuous maps. preprint arXiv:1503.07597
(2015).
[4] Vorono, G. Nouvelles applications des param`etres continus a
` la theorie des formes quadratiques. Deuxi`eme
memoire. Recherches sur les parallello`edres primitifs. J.
Reine Angew. Math. 134, 198287 (1908).
[5] Bernal, J. D. A geometrical approach to the structure of
liquids. Nature (1959).
[6] Rahman, A. Liquid structure and self-diffusion. J. Chem.
Phys. 45, 25852592 (1966).
[7] Finney, J. Random packings and the structure of simple
liquids. Proc. Roy. Soc. Lond. A 319, 479507 (1970).
[8] Bernal, J. D. & Finney, J. L. Random close-packed hardsphere model. II. Geometry of random packing of hard
spheres. Discuss. Faraday Soc. 43, 6269 (1967).
[9] Hsu, C. & Rahman, A. Interaction potentials and their effect on crystal nucleation and symmetry. J. Chem. Phys.
71, 49744986 (1979).
[10] Sheng, H., Luo, W., Alamgir, F., Bai, J. & Ma, E. Atomic
packing and short-to-medium-range order in metallic
glasses. Nature 439, 419425 (2006).
[11] Tanemura, M. et al. Geometrical analysis of crystalliza-

tion of the soft-core model. Prog. Theor. Phys. 58, 1079

1095 (1977).
Lazar, E. A. The Evolution of Cellular Structures
via Curvature Flow. Ph.D. thesis, Princeton University
(2011).
Lazar, E. A., Mason, J. K., MacPherson, R. D.
& Srolovitz, D. J. Complete topology of cells,
grains, and bubbles in three-dimensional microstructures. Phys. Rev. Lett. 109, 95505 (2012).
Weinberg, L. A simple and efficient algorithm for determining isomorphism of planar triply connected graphs.
IEEE Trans. Circuit Theory CT13, 142148 (1966).
Troadec, J., Gervois, A. & Oger, L. Statistics of voronoi
cells of slightly perturbed face-centered cubic and hexagonal close-packed lattices. Europhys. Lett. 42, 167 (1998).
Ackland, G. & Thetford, R. An improved N-body semiempirical model for body-centred cubic transition metals.
Phil. Mag. A 56, 1530 (1987).
Mishin, Y., Mehl, M., Papaconstantopoulos, D., Voter,
A. & Kress, J. Structural stability and lattice defects
in copper: Ab initio, tight-binding, and embedded-atom
calculations. Phys. Rev. B 63, 224106 (2001).
Sun, D. et al. Crystal-melt interfacial free energies in hcp
metals: A molecular dynamics study of Mg. Phys. Rev. B
73, 024116 (2006).
Plimpton, S. Fast parallel algorithms for short-range
molecular dynamics. J. Comput. Phys. 117, 119 (1995).
Einstein, A. Plancks theory of radiation and the theory
of specific heat. Ann. Phys 22, 180190 (1907).
Kawasaki, T. & Tanaka, H. Formation of a crystal nu-

[12]

[13]

[14]

[15]

[16]

[17]

[18]

[19]
[20]
[21]

10

[22]

[23]

[24]

[25]
[26]

[27]

[28]

[29]

[30]

[31]

cleus from liquid. Proc. Natl. Acad. Sci. USA 107, 14036
14041 (2010).
Bai, X.-M. & Li, M. Ring-diffusion mediated homogeneous melting in the superheating regime. Physical Review B 77, 134109 (2008).
Mishin, Y., Farkas, D., Mehl, M. & Papaconstantopoulos, D. Interatomic potentials for monoatomic metals from experimental data and ab initio calculations.
Phys. Rev. B 59, 3393 (1999).
Lazar, E. A., Mason, J. K., MacPherson, R. D. &
Srolovitz, D. J. A more accurate three-dimensional grain
growth algorithm. Acta Mater. 59, 68376847 (2011).
Hirth, J. P. & Lothe, J. Theory of Dislocations (John
Wiley & Sons, 1982).
Foiles, S. M. & Hoyt, J. Computation of grain boundary
stiffness and mobility from boundary fluctuations. Acta
Mater. 54, 33513357 (2006).
Fu, C.-C., Dalla Torre, J., Willaime, F., Bocquet, J.-L.
& Barbu, A. Multiscale modelling of defect kinetics in
irradiated iron. Nature Mater. 4, 6874 (2004).
Kelchner, C. L., Plimpton, S. & Hamilton, J. Dislocation
nucleation and defect structure during surface indentation. Phys. Rev. B 58, 11085 (1998).
Ackland, G. & Jones, A. Applications of local crystal structure measures in experiment and simulation.
Phys. Rev. B 73, 054104 (2006).
Honeycutt, J. D. & Andersen, H. C. Molecular dynamics study of melting and freezing of small Lennard-Jones
clusters. Journal of Physical Chemistry 91, 49504963
(1987).
Kiritani, M. Story of stacking fault tetrahedra. Materials
Chemistry and Physics 50, 133138 (1997).

[32] Stukowski, A. Visualization and analysis of atomistic

simulation data with OVITOthe Open Visualization
Tool. Modelling Simul. Mater. Sci. Eng. 18, 015012
(2010).
[33] Mason, J. K., Lazar, E. A., MacPherson, R. D. &
Srolovitz, D. J. Statistical topology of three-dimensional
poisson-voronoi cells and cell boundary networks. Phys.
Rev. E 88, 063309 (2013).
[34] Krekelberg, W. P., Pond, M. J., Goel, G., Shen, V. K.,
Errington, J. R. & Truskett, T. M. Generalized Rosenfeld scalings for tracer diffusivities in not-so-simple fluids: Mixtures and soft particles. Phys. Rev. E 80, 061205
(2009).
[35] Goel, G., Krekelberg, W. P., Errington, J. R. & Truskett,
T. M. Tuning density profiles and mobility of inhomogeneous fluids. Phys. Rev. Lett. 100, 106001 (2008).
[36] Goel, G., Krekelberg, W. P., Pond, M. J., Mittal, J.,
Shen, V. K., Errington, J. R. & Truskett, T. M. Available
states and available space: static properties that predict
self-diffusivity of confined fluids. J. of Stat. Mech. Theor.
Exp. 04, P04006 (2009).
[37] Stukowski, A., Bulatov, V. V. & Arsenlis, A. Automated
identification and indexing of dislocations in crystal interfaces. Modelling Simul. Mater. Sci. Eng. 20, 085007
(2012).
[38] Rycroft, C. Voro++: A three-dimensional Voronoi cell
library in C++. Chaos 19, 041111 (2009).
[39] Weinberg, L. On the Maximum Order of the Automorphism Group of a Planar Triply Connected Graph. SIAM
J. on Applied Math. 14, 729738 (1966).
[40] Li, J. Atomeye: an efficient atomistic configuration
viewer. Modelling Simul. Mater. Sci. Eng. 11, 173 (2003).

SUPPLEMENTARY MATERIAL

ENUMERATING PRIMARY
AND SECONDARY TYPES

In the paper we considered families of topological types

F associated with particular structures. In this section
we provide some additional detail regarding the determination of these families, and report the numbers of types
in several of them, as well as in the overlap between multiple families.
The Voronoi cell of BCC is topologically stable in the
sense that infinitesimal perturbations of atomic coordinates will not change its topology. Therefore, FBCC consists of a single primary type and no secondary types.
In contrast, Voronoi cells of FCC and HCP are unstable, and infinitesimal perturbations of atomic coordinates will change their topologies. The instability of
these Voronoi cells can be detected in vertices that are
incident with four edges; there are six such unstable vertices in FCC and HCP. As each unstable vertex can either
remain unstable, resolve in one of 3 primary directions,
or resolve in one of 4 secondary directions (see Fig. 3),

we must check 86 = 262,144 configurations that can result from all infinitesimal perturbations. We compute
the topology of each configuration using the algorithm
described in the Materials and Methods section of the
paper. Multiple configurations can result in the same
Voronoi cell topology due to symmetries of the unperturbed configuration.
For FCC we find 44 primary types and 6250 secondary
types; of the secondary types only 2771 have no unstable
vertices. Figure S1 illustrates several Voronoi cells observed in a finite-temperature FCC crystal; their topologies are given by this enumeration technique. For HCP
we find 66 primary types and 21,545 secondary types; of
the secondary types only 9490 have no unstable vertices.
While the determination of topological types associated with a particular structure may require substantial
computation, this needs only be done once per structure. Lookup tables are then created and subsequently
referenced when analyzing atomistic data sets. Data for
several common crystal structures are available from the

11

FIG. S1. Examples of Voronoi cells in a perturbed FCC crystal.

(a)

authors upon request.

In the paper we noted that certain types can belong to
several families (indeterminate types). More specifically,
we note that the unique type belonging to FBCC also belongs to FFCC and FHCP . Furthermore, FFCC and FHCP
share 23 primary types in common and 1352 secondary
types in common.

RESOLVING INDETERMINATE TYPES

In the paper we noted practical challenges created by

the overlap of multiple families of types. In particular,
when attempting to visualize defects in an FCC environment, some atoms were mistakenly identified as belonging to the FCC bulk instead of to the HCP-like defect.
Although a thorough analysis of this problem is beyond
the scope of this paper, we briefly consider one approach
towards resolving it. We leave a complete discussion of
this topic for a future paper.
Indeterminate types are Voronoi topologies that can
be obtained through infinitesimal perturbations of multiple distinct structures. Figure 2 in the paper shows a
number of regions in configuration space incident with
both xFCC and xHCP . One way of deciding whether a
point in this region should be classified as FCC or HCP
involves computing distances in this configuration space:
points close to xFCC should be classified as FCC, while
those close to xHCP should be classified as HCP. Points
closer to xBCC than to either xFCC or xHCP should be
classified as BCC.
While theoretically appealing, there is no practical
manner in which to calculate these distances because
of the complicated topology of this configuration space.
More specifically, the standard metric on R3n cannot be
used due to the actions of the rotation, renormalization,
and permutation groups acting on it.
A more practical approach involves perturbing configurations of particles as follows. If the Voronoi type
of a particle is indeterminate, we randomly perturb the
particle and its nearby neighbors; this corresponds to
a small perturbation of x in configuration space. The
Voronoi cell of the perturbed configuration is calculated;
this procedure is repeated several times. If all perturbations result in indeterminate or HCP-types (the latter
occurring at least once), then the particle is classified as
being HCP-type.
We consider a stacking-fault tetrahedron (SFT) in a

(b)

FIG. S2. A high-temperature single crystal containing an

SFT, visualized using (a) the method described in the paper
and (b) the modified method proposed here.

single copper crystal that was heated to 85% of its melting temperature; this system was considered in Comparison with Other Methods section of the paper. Figure
S2 shows the SFT colored by the approach considered in
the paper and by the modification considered here. This
modified approach provides a more robust visualization
of the SFT than the approach suggested in the paper.
We defer a complete discussion of indeterminate types
and methods of resolving them to a future paper.

QUANTITATIVE COMPARISON
AT HIGH TEMPERATURES

In the paper we have shown that Voronoi topology enables robust visualization of structure in hightemperature systems that cannot be obtained using conventional methods. Here we provide a direct and quantitative comparison beyond visual inspection. We begin with a system containing 1,372,000 copper atoms
organized in a perfect crystal, and heat the system to
just below its melting point as described in the FiniteTemperature Crystals section of the paper. For each
order-parameter considered in the paper, we calculate the
frequency of atoms in this single crystal that are characterized as non-FCC-type as a function of temperature.
Close inspection of the system shows that there are no
defects.
For each visualization method we considering the number of atoms in this system characterized as non-FCCtype. We use OVITO to compute bond-angle analysis,
adaptive common-neighbor analysis, and centrosymmetry values for each atom and at each temperature. Figure S3 shows the frequency of non-FCC-type atoms, as
measured by each of the order-parameters, as a function
of temperature.
In contrast to the bond-angle and adaptive commonneighbor classifications provided by OVITO, centrosymmetry (CS) is reported as a scalar requiring choice of a
cutoff; atoms with CS values above this cutoff are not
considered FCC. To determine an appropriate cutoff we
considered CS values of atoms adjacent to either a vacancy or an interstitial in an otherwise perfect FCC crys-

12

Frequency of non-FCC-type atoms

0 K, colored using the same order-parameters considered

in the paper.
Although all methods successfully detect the presence of the SFT in the quenched system, centrosymmetry, bond-angle analysis, and adaptive common-neighbor
analysis also identify atoms outside the SFT as being
non-FCC-type. These non-FCC-type identifications are
the result of small local strains which are not structural
defects. Visualization of the sample through Voronoi
topology does not result in any such misidentifications,
providing the most robust picture of structural defects.

Adaptive CNA
Bond-Angle Analysis
Centrosymmetry > 2
Centrosymmetry > 6
Voronoi Topology

0.8

0.6

0.4

0.2

200

400

600

800

1000

1200

Temperature (K)

FIG. S3. Frequency of atoms characterized as non-FCC-type

in a copper crystal.

tal. Atoms adjacent to a vacancy have a CS value of 6.20;

interstitial atoms have a CS value of 13.88. In order that
our choice of cutoff allows for the detection of vacancies,
we choose a cutoff of 6. We note that cutoff values as low
as 0.5 and 1 can be regularly found in the literature, and
so our choice of cutoff is extremely conservative.
At T = 1300 K, bond-angle analysis, adaptive
common-neighbor analysis, and centrosymmetry incorrectly identify over 36.9%, 55.7% and 13.5%, respectively,
of the atoms as not belonging to the FCC bulk crystal.
By contrast, Voronoi topology mistakenly identifies only
0.53% as such atoms as not belonging to the FCC bulk
crystal.

QUENCHED SYSTEMS

The topological approach is not limited to studying

finite-temperature systems, and can also be applied to
structures obtained through quenching high-temperature
samples. Figure S4 shows the same stacking-fault tetrahedron (SFT) considered in the paper after quenching to

(a) Centrosymmetry

(b) Bond-angle analysis

(c) Adaptive CNA

(d) Voronoi Topology

FIG. S4. Cross-section of a stacking-fault tetrahedron in copper, after quenching from 85% of its melting temperature,
colored using several popular visualization approaches, and
Voronoi topology. In (b) and (c), dark blue, yellow and red
indicate atoms in FCC, HCP, and other local environments,
respectively. In (d), dark blue, yellow, and red indicate atoms
that are FCC-types, HCP- but not FCC-types, and all other
types, respectively.